Responsible AI Principles

A university deploys an AI-powered admissions system and tests it to make sure every application is judged on relevant academic criteria, avoiding unfounded discrimination based on irrelevant demographic factors. Which responsible AI principle is being applied?

• AAccountability
• BTransparency
• CInclusiveness
• DFairness✓

Answer: Fairness requires AI developers to minimise bias in training data and to test AI systems so that they do not produce discriminatory outputs. Checking that an admissions model judges applications on academic merit rather than on irrelevant demographic factors is Microsoft's own example of that test. Transparency is about telling users how a system works, inclusiveness is about not excluding users from the benefits of AI, and accountability is about governance of the people and organisations that build it.

A warehouse robot uses computer vision to identify objects before picking them up. To avoid unintentional damage, its developers make it take no action when the model's confidence value for an object falls below a set threshold. Which responsible AI principle does this practice support?

• AReliability and safety✓
• BPrivacy and security
• CInclusiveness
• DAccountability

Answer: Reliability and safety recognises that AI is based on probabilistic models and is not infallible, so AI-powered applications must mitigate the risks accordingly. Using probability values as a confidence measure and refusing to act below a threshold is Microsoft's example of that mitigation for a robot that interacts with physical objects. Privacy and security concerns personal data, accountability concerns governance, and inclusiveness concerns not excluding users.

An airport uses a facial identification system to grant travellers temporary access to a secure area. Which practice does the privacy and security principle call for in this scenario?

• AOffer an alternative non-visual identification method so that no traveller is excluded from using it
• BTest the system to ensure it does not discriminate against travellers from any demographic group
• CDelete the personal images as soon as they are no longer required and restrict who can view them✓
• DPublish a description of the features of its training data so that travellers understand how it works

Answer: Privacy and security requires that personal data handled by an AI system is kept secure and cannot be revealed to people with no need to see it. Microsoft's airport example deletes the images used for temporary access as soon as they are no longer required and adds safeguards so that operators or users without a need to view them cannot access them. Describing the training data relates to transparency, discrimination testing to fairness, and alternative access methods to inclusiveness.

A company builds an AI agent that users interact with entirely by voice. So that people with a hearing impairment can still use it, the team makes the agent generate text captions as well. Which responsible AI principle does this address?

• AAccountability
• BFairness
• CInclusiveness✓
• DTransparency

Answer: Inclusiveness means the potential of AI to improve lives and drive success should be open to everyone, so developers should strive to ensure their solutions do not exclude some users. Generating text captions for a speech-based agent is Microsoft's example of keeping the system usable for people with a hearing impairment. Transparency concerns explaining how the system works, accountability concerns governance, and fairness concerns avoiding discriminatory outputs caused by biased data.

A bank uses an AI-based loan-approval application. It discloses that AI is used and describes the features of the data on which the model was trained, without revealing confidential information. Which responsible AI principle is the bank applying?

• AFairness
• BTransparency✓
• CInclusiveness
• DAccountability

Answer: Transparency means making users aware of how an AI system works and any potential limitations it may have, because AI can otherwise seem like magic. Disclosing the use of AI and describing the features of the training data is Microsoft's example for a loan-approval application. Accountability is about the governance framework of the organisation, inclusiveness is about not excluding users, and fairness is about avoiding discriminatory outputs.

The leadership of an organisation that develops and distributes AI applications defines and applies a framework of governance to make sure responsible AI principles are followed on every project. Which principle does this most directly fulfil?

• AAccountability✓
• BTransparency
• CInclusiveness
• DReliability and safety

Answer: Accountability states that the people and organisations who develop and distribute AI solutions are ultimately accountable for their actions. Microsoft advises organisations developing AI models and applications to define and apply a framework of governance to help ensure they apply responsible AI principles to their work. Transparency is about informing users, reliability and safety is about mitigating the risks of probabilistic models, and inclusiveness is about not excluding users.

A product team believes that enabling content filters on its generative AI chat application fully satisfies responsible AI. Which statement best describes the role of content filters?

• AThe complete responsible AI solution, so no further design or operational consideration is needed once they are switched on
• BOne way to mitigate the risk of harmful content, but not a substitute for applying the principles from conception to operation✓
• CA replacement for a governance framework, because filtered outputs cannot produce harmful, illegal or offensive material
• DA control applied only to the training data, so they have no effect on the content that a deployed model generates at run time

Answer: Content filters are one way that AI systems mitigate the risk of harmful content generation. A responsible AI solution still requires consideration of the key principles from its conception, through its design and implementation, and into its operation. Filters are therefore neither the whole solution nor a replacement for governance, and they act on generated content rather than only on training data.

A data science team is about to train a model that will influence hiring recommendations. What does the fairness principle require of them?

• ADefine and apply a governance framework covering the model's use
• BMake users aware of how the system works and any limitations it has
• CKeep the training data secure and ensure the model cannot reveal private details
• DMinimise bias in the training data and test the system for fairness✓

Answer: Fairness addresses the substantial risk that data selection criteria, or the data itself, reflect unconscious bias that causes a model to produce discriminatory outputs. AI developers therefore need to take care to minimise bias in training data and to test AI systems for fairness. Securing data and preventing disclosure belongs to privacy and security, informing users belongs to transparency, and governance belongs to accountability.

Which characteristic of AI is the reason the reliability and safety principle exists?

• AAI models may be trained on data that includes personal information
• BAI can seem like magic to users who do not know how it works
• CAI is based on probabilistic models and is not infallible✓
• DAI is trained on data that is sourced and selected by humans

Answer: AI is based on probabilistic models, so it is not infallible, and AI-powered applications need to take this into account and mitigate risks accordingly. That is the reasoning behind the reliability and safety principle. Human-selected data is the basis of the fairness principle, personal information in training data underpins privacy and security, and the sense that AI seems like magic motivates transparency.

A team runs Foundry safety evaluators against the outputs of a customer-facing agent. What should they expect the evaluators to do?

• AApply content filters to the training data so the model cannot learn harmful content
• BRewrite, redact or block unsafe responses automatically before they reach the user
• CBlock the deployment until every flagged safety issue has been resolved
• DDetect, scan and score harmful or unsafe content without actively resolving it✓

Answer: Foundry evaluators are components that measure the quality, safety and effectiveness of model or agent outputs, and safety evaluators scan for harmful or unsafe content, bias and unfairness, violence, self-harm and protected-class harms. On their own the evaluators detect, scan and score issues but do not actively resolve them. They do not rewrite responses, gate the deployment or filter training data.

A healthcare provider fine-tunes a model on patient records and stores the training data securely. Under the privacy and security principle, what further responsibility do the developers have?

• AEnsure a governance framework assigns clear accountability for every recommendation the model produces in practice
• BEnsure the model's outputs are tested for discriminatory bias across every patient group it serves
• CEnsure the trained model itself cannot be used to reveal private personal or organisational details✓
• DEnsure clinicians are made aware of how the model works and the limitations it has

Answer: Privacy and security places two responsibilities on AI developers: keeping the training data secure, and ensuring that the trained models themselves cannot be used to reveal private personal or organisational details. Securing the data alone therefore does not discharge the principle. Bias testing belongs to fairness, governance to accountability, and explaining the model's workings and limitations to transparency.

Before selecting a model from the Foundry model catalog, an architect wants to understand the model's known constraints and caveats. Which part of the catalog entry should she read?

• AThe list of supported inference tasks and fine-tuning options
• BThe benchmark results, leaderboards and performance comparisons
• CThe tokens per minute allocation and deployment type settings
• DThe Responsible AI documentation, such as the model card✓

Answer: Each model entry in the Foundry model catalog typically includes Responsible AI documentation, described as model cards, constraints and caveats, which is where a model's known limitations are set out for anyone evaluating it. Benchmark results compare performance, the inference-task list shows what the model supports, and TPM and deployment type are configured when a model is deployed rather than read from its catalog entry.